Home Page

Approach

Datasets

• The TDAmeritrade API was resposnible for compiling basic finacial data for $AAPL and a number of comparative assests over the past 10 years including:
  • Open
  • Close
  • High
  • Low
  • Volume
• Comparative Assets
  • To accurately gauge and incorporate industry and macro trends, we have compiled the basic technical indicators of competitive companies and industry indexes. These included:
    • $META
    • $GOOGL
    • $MSFT
    • $AMZN
    • $SP500
• Tweets mentioning “$AAPL” and “AAPL” are scraped from Twitter.
• The dataset includes data from 05/09/2013 to 03/31/2023

Technical Analysis

Feature Creation

• Technical indicators are computed from our initial set of basic indicators (Open, Close, High, Low, and Volume). Some of the technical indicators computed are:
  • Stochastic Oscillators
  • Relative Strength Index
  • Simple Moving Averages for Close and Volume
  • Moving Average Convergence/Divergence
• Below demonstrates an example of the use of technical indicators over the last 400 days by showing the correlation of:
  • SMA 7 (Green)
  • SMA 21 (Red)
  • Upper Band + Lower Band (Cyan)
  • Closing Price of $AAPL (Dark Blue)

Feature Extraction

(Note: While attempted, autoencoding did not offer tangible model improvements)

• An autoencoder is used to produce a set of compressed features from the combination of basic and technical indicators.
• The compressed representation of the full technical indicator set is fed into the prediction model for training and forecasting purposes.

Fundamental Analysis

• The usage of sentiment analysis of data coming directly from the public (social media websites), rather than from biased news outlets or similar feeds.
• Text content from Twitter are web-scraped and processed via the NLTK library.
• VADER is used upon each media post to compute sentiment scores which are then averaged for any given day.
• A polarity score between -1 and 1 is assigned for average sentiment on a given day.
• Below demonstrates the recorded sentiment values of $AAPL stock over the duration of our dataset.

Prediction Model

• After feature engineering, our final set of compressed technical indicators is appended with the sentiment scores from the fundamental analysis and fed into a wGAN-GP model for training.
• wGAN-GP Model Description
  • Wasserstein distance is used as a loss function for promotion of training stability.
  • Gradient norm penalty is enforced in the discriminator in hopes of achieving Lipschitz continuity.
  • Generator is an LSTM with input units equal to the final number of features and 512 hidden units, leading into a final linear layer with a single output for the closing price of each day.
  • Discriminator is a 13-Layer CNN composed of 1-D Convolutional, LeakyReLU, and Dense layers.
• Model is tuned with a predefined grid search.

Generator

Discriminator